Applied Business Forecasting and Planning The BoxJenkins Methodology

Applied Business Forecasting and Planning The Box-Jenkins Methodology for ARIMA Models

Introduction n Autoregressive Integrated Moving Average models (ARIMA models) were popularized by George Box and Gwilym Jenkins in the early 1970 s. ARIMA models are a class of linear models that is capable of representing stationary as well as nonstationary time series. ARIMA models do not involve independent variables in their construction. They make use of the information in the series itself to generate forecasts.

Introduction n ARIMA models rely heavily on autocorrelation patterns in the data. ARIMA methodology of forecasting is different from most methods because it does not assume any particular pattern in the historical data of the series to be forecast. It uses an interactive approach of identifying a possible model from a general class of models. The chosen model is then checked against the historical data to see if it accurately describe the series.

Introduction n Recall that, a time series data is a sequence of numerical observations naturally ordered in time n n n Daily closing price of IBM stock Weekly automobile production by the Pontiac division of general Motors. Hourly temperatures at the entrance to Grand central Station.

Introduction n Two question of paramount importance When a forecaster examines a time series data are: n n Do the data exhibit a discernible pattern? Can this be exploited to make meaningful forecasts?

Introduction n n The Box-Jenkins methodology refers to a set of procedures for identifying, fitting, and checking ARIMA models with time series data. Forecasts follow directly from the form of fitted model. The basis of BOX-Jenkins approach to modeling time series consists of three phases: n n n Identification Estimation and testing Application

Introduction n Identification n Data preparation n Transform data to stabilize variance Differencing data to obtain stationary series Model selection n Examine data, ACF and PACF to identify potential models

Introduction n Estimation and testing n Estimation n Estimate parameters in potential models Select best model using suitable criterion Diagnostics n n n Check ACF/PACF of residuals Do portmanteau test of residuals Are the residuals white noise?

Introduction n Application n Forecasting: use model to forecast

Examining correlation in time series data n n The key statistic in time series analysis is the autocorrelation coefficient ( the correlation of the time series with itself, lagged 1, 2, or more periods. ) Recall the autocorrelation formula:

Examining Correlation in Time Series Data n n n Recall r 1 indicates how successive values of Y relate to each other, r 2 indicates how Y values two periods apart relate to each other, and so on. The auto correlations at lag 1, 2, …, make up the autocorrelation function or ACF. Autocorrelation function is a valuable tool for investigating properties of an empirical time series.

A white noise model n n n A white noise model is a model where observations Yt is made of two parts: a fixed value and an uncorrelated random error component. For uncorrelated data (a time series which is white noise) we expect each autocorrelation to be close to zero. Consider the following white noise series.

White noise series

ACF for the white noise series

Sampling distribution of autocorrelation n n The autocorrelation coefficients of white noise data have a sampling distribution that can be approximated by a normal distribution with mean zero and standard error 1/ n. where n is the number of observations in the series. This information can be used to develop tests of hypotheses and confidence intervals for ACF.

Sampling distribution of autocorrelation n For example n n n For our white noise series example, we expect 95% of all sample ACF to be within If this is not the case then the series is not white noise. The sampling distribution and standard error allow us to distinguish what is randomness or white noise from what is pattern.

Portmanteau tests n n Instead of studying the ACF value one at a time, we can consider a set of them together, for example the first 10 of them (r 1 through r 10) all at one time. A common test is the Box-Pierce test which is based on the Box-Pierce Q statistics n Usually h 20 is selected

Portmanteau tests n n This test was originally developed by Box and Pierce for testing the residuals from a forecast model. Any good forecast model should have forecast errors which follow a white noise model. If the series is white noise then, the Q statistic has a chi-square distribution with (h-m) degrees of freedom, where m is the number of parameters in the model which has been fitted to the data. The test can easily be applied to raw data, when no model has been fitted , by setting m = 0.

Example n Here is the ACF values for the white noise example.

Example n n The box-Pierce Q statistics for h = 10 is Since the data is not modeled m =0 therefore df = 10. From table C-4 with 10 df, the probability of obtaining a chi-square value as large or larger than 5. 66 is greater than 0. 1. The set of 10 rk values are not significantly different from zero.

Portmanteau tests n n n An alternative portmanteau test is the Ljung-Box test. Q* has a Chi-square distribution with (h-m) degrees of freedom. In general, the data are not white noise if the values of Q or Q* is greater than the value given in a chi square table with = 5%.

The Partial autocorrelation coefficient n n Partial autocorrelations measures the degree of association between yt and yt-k, when the effects of other time lags 1, 2, 3, …, k-1 are removed. The partial autocorrelation coefficient of order k is evaluated by regressing yt against yt-1, …yt-k: n k (partial autocorrelation coefficient of order k) is the estimated coefficient bk.

The Partial autocorrelation coefficient n n n The partial autocorrelation functions (PACF) should all be close to zero for a white noise series. If the time series is white noise, the estimated PACF are approximately independent and normally distributed with a standard error 1/ n. Therefore the same critical values of Can be used with PACF to asses if the data are white noise.

The Partial autocorrelation coefficient n n It is usual to plot the partial autocorrelation function or PACF. The PACF plot of the white noise data is presented in the next slide.

PACF plot of the white noise series.

Examining stationarity of time series data n n n Stationarity means no growth or decline. Data fluctuates around a constant mean independent of time and variance of the fluctuation remains constant over time. Stationarity can be assessed using a time series plot. n n Plot shows no change in the mean over time No obvious change in the variance over time.

Examining stationarity of time series data n The autocorrelation plot can also show nonstationarity. n n Significant autocorrelation for several time lags and slow decline in rk indicate nonstationarity. The following graph shows the seasonally adjusted sales for Gap stores from 1985 to 2003.

Examining stationarity of time series data

Examining stationarity of time series data n n The time series plot shows that it is nonstationary in the mean. The next slide shows the ACF plot for this data series.

Examining stationarity of time series data

Examining stationarity of time series data n The ACF also shows a pattern typical for a non-stationary series: n n n Large significant ACF for the first 7 time lag Slow decrease in the size of the autocorrelations. The PACF is shown in the next slide.

Examining stationarity of time series data

Examining stationarity of time series data n This is also typical of a non-stationary series. n Partial autocorrelation at time lag 1 is close to one and the partial autocorrelation for the time lag 2 through 18 are close to zero.

Removing non-stationarity in time series n n The non-stationary pattern in a time series data needs to be removed in order that other correlation structure present in the series can be seen before proceeding with model building. One way of removing non-stationarity is through the method of differencing.

Removing non-stationarity in time series n n The differenced series is defined as: The following two slides shows the time series plot and the ACF plot of the monthly S&P 500 composite index from 1979 to 1997.

Removing non-stationarity in time series

Removing non-stationarity in time series n n n The time plot shows that it is not stationary in the mean. The ACF and PACF plot also display a pattern typical for non-stationary pattern. Taking the first difference of the S& P 500 composite index data represents the monthly changes in the S&P 500 composite index.

Removing non-stationarity in time series n n The time series plot and the ACF and PACF plots indicate that the first difference has removed the growth in the time series data. The series looks just like a white noise with almost no autocorrelation or partial autocorrelation outside the 95% limits.

Removing non-stationarity in time series

Removing non-stationarity in time series n Note that the ACF and PACF at lag 1 is outside the limits, but it is acceptable to have about 5% of spikes fall a short distance beyond the limit due to chance.

Random Walk n n Let yt denote the S&P 500 composite index, then the time series plot of differenced S&P 500 composite index suggests that a suitable model for the data might be Where et is white noise.

Random Walk n n The equation in the previous slide can be rewritten as This model is known as “random walk” model and it is widely used for nonstationary data.

Random Walk n n Random walks typically have long periods of apparent trends up or down which can suddenly change direction unpredictably They are commonly used in analyzing economic and stock price series.

Removing non-stationarity in time series n Taking first differencing is a very useful tool for removing non-statioanarity, but sometimes the differenced data will not appear stationary and it may be necessary to difference the data a second time.

Removing non-stationarity in time series n n The series of second order difference is defined: In practice, it is almost never necessary to go beyond second order differences.

Seasonal differencing n n With seasonal data which is not stationary, it is appropriate to take seasonal differences. A seasonal difference is the difference between an observation and the corresponding observation from the previous year. n Where s is the length of the season

Seasonal differencing n n n The Gap quarterly sales is an example of a non-stationary seasonal data. The following time series plot show a trend with a pronounced seasonal component The auto correlations show that n n The series is non-stationary. The series is seasonal.

Seasonal differencing

Seasonal differencing n n The seasonally differenced series represents the change in sales between quarters of consecutive years. The time series plot, ACF and PACF of the seasonally differenced Gap’s quarterly sales are in the following three slides.

Seasonal differencing

Seasonal differencing n n The series is now much closer to being stationary, but more than 5% of the spikes are beyond 95% critical limits and autocorrelation show gradual decline in values. The seasonality is still present as shown by spike at time lag 4 in the PACF.

Seasonal differencing n n The remaining non-stationarity in the mean can be removed with a further first difference. When both seasonal and first differences are applied, it does not make no difference which is done first.

Seasonal differencing n n It is recommended to do the seasonal differencing first since sometimes the resulting series will be stationary and hence no need for a further first difference. When differencing is used, it is important that the differences be interpretable.

Seasonal differencing n n The series resulted from first difference of seasonally differenced Gap’s quarterly sales data is reported in the following three slides. Is the resulting series white noise?

Seasonal differencing

Tests for stationarity n n n Several statistical tests has been developed to determine if a series is stationary. These tests are also known as unit root tests. One of the widely used such test is the Dickey-fuller test.

Tests for stationarity n n To carry out the test, fit the regression model Where The number of lagged terms p, is usually set to 3.

Tests for stationarity n n n The value of is estimated using ordinary least squares. If the original series yt needs differencing, the estimated value of will be close to zero. If yt is already stationary, the estimated value of will be negative.

ARIMA models for time series data n Autoregression n Consider regression models of the form n Define

ARIMA models for time series data n n n Then the previous equation becomes The explanatory variables in this equations are time-lagged values of the variable y. Autoregression (AR) is used to describe models of this form.

ARIMA models for time series data n Autoregression models should be treated differently from ordinary regression models since: n n The explanatory variables in the autoregression models have a built-in dependence relationship. Determining the number of past values of yt to include in the model is not always straight forward.

ARIMA models for time series data n Moving average model n A time series model which uses past errors as explanatory variable: is called moving average(MA) model n Note that this model is defined as a moving average of the error series, while the moving average models we discussed previously are the moving average of the observations.

ARIMA models for time series data n n Autoregressive (AR) models can be coupled with moving average (MA) models to form a general and useful class of time series models called Autoregressive Moving Average (ARMA) models. These can be used when the data are stationary.

ARIMA models for time series data n n n This class of models can be extended to non -stationary series by allowing the differencing of the data series. These are called Autoregressive Integrated Moving Average(ARIMA) models. There a large variety of ARIMA models.

ARIMA models for time series data n The general non-seasonal model is known as ARIMA (p, d, q): n n n p is the number of autoregressive terms. d is the number of differences. q is the number of moving average terms.

ARIMA models for time series data n A white noise model is classified as ARIMA (0, 0, 0) n n n No AR part since yt does not depend on yt-1. There is no differencing involved. No MA part since yt does not depend on et-1.

ARIMA models for time series data n A random walk model is classified as ARIMA (0, 1, 0) n n n There is no AR part. There is no MA part. There is one difference.

ARIMA models for time series data n n Note that if any of p, d, or q are equal to zero, the model can be written in a shorthand notation by dropping the unused part. Example n n ARIMA(2, 0, 0) = AR(2) ARIMA (1, 0, 1) = ARMA(1, 1)

An autoregressive model of order one AR(1) n The basic form of an ARIMA (1, 0, 0) or AR(1) is: n n Observation yt depends on y t-1. The value of autoregressive coefficient 1 is between – 1 and 1.

An autoregressive model of order one n The time plot of an AR(1) model varies with the parameter 1. . n n When 1= 0, yt is equivalent to a white noise series. When 1= 1, yt is equivalent to a random walk series For negative values of 1, the series tends to oscillate between positive and negative values. The following slides show the time series, ACF and PACF plot for an ARIMA(1, 0, 0) time series data.

An autoregressive model of order one

An autoregressive model of order one n The ACF and PACF can be used to identify an AR(1) model. n n The autocorrelations decay exponentially. There is a single significant partial autocorrelation.

A moving average of order one MA(1) n The general form of ARIMA (0, 0, 1) or MA(1) model is n n n Yt depends on the error term et and on the previous error term et-1 with coefficient - 1. The value of 1 is between – 1 and 1. The following slides show an example of an MA(1) data series.

A moving average of order one MA(1)

A moving average of order one MA(1) n n Note that there is only one significant autocorrelation at time lag 1. The partial autocorrelations decay exponentially, but because of random error components, they do not die out to zero as do theoretical autocorrelation.

Higher order auto regressive models n A pth-order AR model is defined as n n n C is the constant term j is the jth auto regression parameter et is the error term at time t.

Higher order auto regressive models n Restrictions on the allowable values of auto regression parameters n For p =1 n n -1< 1 For p = 2 n n n -1< 2 < 1 1+ 2 <1 2 - 1 <1

Higher order auto regressive models n n A great variety of time series are possible with autoregressive models. The following slides shows an AR(2) model. Note that for AR(2) models the autocorrelations die out in a damped Sine-wave patterns. There are exactly two significant partial autocorrelations.

Higher order auto regressive models

Higher order moving average models n The general MA model of order q can be written as n n n C is the constant term j is the jth moving average parameter. e t-k is the error term at time t-k

Higher order moving average models n Restrictions on the allowable values of the MA parameters. n For q =1 n n -1 < 1 For q =2 n n n -1 < 2 < 1 1 + 2 < 1 2 - 1 < 1

Higher order moving average models n n n A wide variety of time series can be produced using moving average models. In general, the autocorrelations of an MA(q) models are zero beyond lag q For q 2, the PACF can show exponential decay or damped sine-wave patterns.

Mixtures ARMA models n n Basic elements of AR and MA models can be combined to produce a great variety of models. The following is the combination of MA(1) and AR(1) models This is model called ARMA(1, 1) or ARIMA (1, 0, 1) The series is assumed stationary in the mean and in the variance.

Mixtures ARIMA models n n If non-stationarity is added to a mixed ARMA model, then the general ARIMA (p, d, q) is obtained. The equation for the simplest ARIMA (1, 1, 1) is given below.

Mixtures ARIMA models n n n The general ARIMA (p, d, q) model gives a tremendous variety of patterns in the ACF and PACF, so it is not practical to state rules for identifying general ARIMA models. In practice, it is seldom necessary to deal with values p, d, or q that are larger than 0, 1, or 2. It is remarkable that such a small range of values for p, d, or q can cover such a large range of practical forecasting situations.

Seasonality and ARIMA models n n The ARIMA models can be extended to handle seasonal components of a data series. The general shorthand notation is ARIMA (p, d, q)(P, D, Q)s n Where s is the number of periods per season.

Seasonality and ARIMA models n n The general ARIMA(1, 1, 1)4 can be written as Once the coefficients 1, Ф 1, θ 1, and 1 have been estimated from the data, the above equation can be used forecasting.

Seasonality and ARIMA models n n The seasonal lags of the ACF and PACF plots show the seasonal parts of an AR or MA model. Examples: n Seasonal MA model: n ARIMA(0, 0, 0)(0, 0, 1)12 n n will show a spike at lag 12 in the ACF but no other significant spikes. The PACF will show exponential decay in the seasonal lags i. e. at lags 12, 24, 36, …

Seasonality and ARIMA models Seasonal AR model: n ARIMA(0, 0, 0)(1, 0, 0)12 n n will show exponential decay in seasonal lags of the ACF. Single significant spike at lag 12 in the PACF.

Implementing the model –Building Strategy n The Box –Jenkins approach uses an iterative model-building strategy that consist of n n n Selecting an initial model (model identification) Estimating the model coefficients (parameter estimation) Analyzing the residuals (model checking)

Implementing the model –Building Strategy n If necessary, the initial model is modified and the process is repeated until the residual indicate no further modification is necessary. At this point the fitted model can be used forecasting.

Model identification n The following approach outlines an approach to select an appropriate model among a large variety of ARIMA models possible. n Plot the data n n Identify any unusual observations If necessary, transform the dat to stabilize the variance

Model identification n n Check the time series plot, ACF, PACF of the data (possibly transformed) for stationarity. IF n n n Time plot shows the data scattered horizontally around a constant mean ACF and PACF to or near zero quickly Then, the data are stationary.

Model identification n Use differencing to transform the data into a stationary series n n n For no-seasonal data take first differences For seasonal data take seasonal differences Check the plots again if they appear nonstationary, take the differences of the differenced data.

Model identification When the stationarity has been achieved, check the ACF and PACF plots for any pattern remaining. There are three possibilities n n n AR or MA models n n No significant ACF after time lag q indicates MA(q) may be appropriate. No significant PACF after time lag p indicates that AR(p) may be appropriate.

Model identification n n Seasonality is present if ACF and/or PACF at the seasonal lags are large and significant. If no clear MA or AR model is suggested, a mixture model may be appropriate.

Model identification n Example n Non seasonal time series data. n n The following example looks at the number of users logged onto an internet server over a 100 minutes period. The time plot, ACF and PACF is reported in the following three slides.

Model identification

Model identification n n The gradual decline of ACF values indicates nonstationary series. The first partial autocorrelation is very dominant and close to 1, indicating non-stationarity. The time series plot clearly indicates nonstationarity. We take the first differences of the data and reanalyze.

Model identification

Model identification n n ACF shows a mixture of exponential decay and sine-wave pattern PACF shows three significant PACF values. This suggests an AR(3) model. This identifies an ARIMA(3, 1, 0).

Model identification n Example n A seasonal time series. n n n The following example looks at the monthly industry sales (in thousands of francs) for printing and writing papers between the years 1963 and 1972. The time plot, ACF and PACF shows a clear seasonal pattern in the data. This is clear in the large values at time lag 12, 24 and 36.

Model identification

Model identification n We take a seasonal difference and check the time plot, ACF and PACF. The seasonally differenced data appears to be non-stationary (the plots are not shown), so we difference the data again. the following three slides show the twice differenced series.

Model identification

Model identification n n The PACF shows the exponential decay in values. The ACF shows a significant value at time lag 1. n n This suggest a MA(1) model. The ACF also shows a significant value at time lag 12 n This suggest a seasonal MA(1).

Model identification n Therefore, the identifies model is ARIMA (0, 1, 1)12. This model is sometimes is called the “airline model” because it was applied to international airline data by Box and Jenkins. It is one of the most commonly used seasonal ARIMA model.

Model identification n Example 3 n A seasonal data needing transformation n n In this example we look at the monthly shipments of a company that manufactures pollution equipments The time plot shows that the variability increases as the time increases. This indicate that the data is nonstationary in the variance.

Model identification

Model identification n We need to stabilize the variance before fitting an ARIMA model. Logarithmic or power transformation of the data will make the variance stationary. The time plot, ACF and PACF for the logged data is reported in the following three slides.

Model identification

Model identification n n The time plot shows that the magnitude of the fluctuations in the log-transformed data does not vary with time. But, the logged data are clearly non-stationary. n n n The gradual decay of the ACF values. To achieve stationarity, we take the first differences of the logged data. The plots are reported in the next three slides.

Model identification

Model identification n There are significant spikes at time lag 1 and 2 in the PACF, indicating an AR(2) might be appropriate. The single significant spike at lag 12 of the PACF indicates a seasonal AR(1) component. Therefore for the logged data a tentative model would be ARIMA(2, 1, 0)(1, 0, 0)12

Summary n The process of identifying an ARIMA model requires experience and good judgment. The following guidelines can be helpful. n Make the series stationary in mean and variance n n Differencing will take care of non-stationarity in the mean. Logarithmic or power transformation will often take care of non-stationarity in the variance.

Summary n Consider non-seasonal aspect n The ACF and PACF of the stationary data obtained from the previous step can reveal whether MA of AR is feasible. n n Exponential decay or damped sine-wave. For ACF, and spikes at lags 1 to p then cut off to zero, indicate an AR(P) model. Spikes at lag 1 to q, then cut off to zero for ACF and exponential decay or damped sine-wave for PACF indicates MA(q) model.

Summary n Consider seasonal aspect n Examination of ACF and PACF at the seasonal lags can help to identify AR and MA models for the seasonal aspect of the data. n For example, for quarterly data the pattern of r 4, r 8, r 12, r 16, and so on.

Backshift notation n Backward shift operator, B, is defined as Two applications of B to Yt, shifts the data back two periods: A shift to the same quarter last year will use B 4 which is

Backshift notation n n The backward shift operator can be used to describe the differencing process. A first difference can be written as The second order differences as

Backshift notation n Example; n ARMA(1, 1) or ARIMA(1, 0, 1) model n ARMA(p, q) or ARIMA(p, 0, q) model

Backshift notation n ARIMA(1, 1, 1)

Estimating the parameters n n Once a tentative model has been selected, the parameters for the model must be estimated. The method of least squares can be used for RIMA model. However, for models with an MA components, there is no simple formula that can be used to estimate the parameters. Instead, an iterative method is used. This involves starting with a preliminary estimate, and refining the estimate iteratively until the sum of the squared errors is minimized.

Estimating the parameters n n n Another method of estimating the parameters is the maximum likelihood procedure. Like least squares methods, these estimates must be found iteratively. Maximum likelihood estimation is usually favored because it has some desiable statistical properties.

Estimating the parameters n n After the estimates and their standard errors are determined, t values can be constructed and interpreted in the usual way. Parameters that are judged significantly different from zero are retained in the fitted model; parameters that are not significantly different from zero are dropped from the model.

Estimating the parameters n n There may have been more than one plausible model identified, and we need a method to determine which of them is preferred. Akaike’s Information Criterion (AIC) n n L denotes the likelihood m is the number of parameters estimated in the model: m = p+q+P+Q

Estimating the parameters n n Because not all computer programs produce the AIC or the likelihood L, it is not always possible to find the AIC for a given model. A useful approximation to the AIC is:

Diagnostic Checking n n n Before using the model forecasting, it must be checked for adequacy. A model is adequate if the residuals left over after fitting the model is simply white noise. The pattern of ACF and PACF of the residuals may suggest how the model can be improved.

Diagnostic Checking n For example n n Significant spikes at the seasonal lags suggests adding seasonal component to the chosen model Significant spikes at small lags suggest increasing the non-seasonal AR or MA components of the model.

Diagnostic Checking n n A portmanteau test can also be applied to the residuals as an additional test of fit. If the portmanteau test is significant, then the model is inadequate. In this case we need to go back and consider other ARIMA models. Any new model will need their parameters estimated and their AIC values computed and compared with other models.

Diagnostic Checking n n Usually, the model with the smallest AIC will have residuals which resemble white noise. Occasionally, it might be necessary to adopt a model with not quite the smallest AIC value, but with better behaved residuals.

Example n The analyst for the ISC Corporation was asked to develop forecasts for the closing prices of ISC stock. The stock has been languishing for some time with little growth, and senior management wanted some projections to discuss with the board of directors. The ISC stock prices are plotted in the following slide.

Example

Example n n The plot of the stock prices suggests the series is stationary. The stock prices vary about a fixed level of approximately 250. Is the Box-Jenkins methodology appropriate for this data series? The ACF and PACF for the stock price series are reported in the following two slides.

Example

Example n n n The sample ACF alternate in sign and decline to zero after lag 2. The sample PACF are similar are close to zero after time lag 2. These are consistent with an AR(2) or ARIMA(2, 0, 0) model Using MINITAB an AR(2) model is fit to the data. WE include a constant term to allow for a nonzero level.

Example n n n The estimated coefficient 2 is not significant (t=1. 75) at 5% level but is significant at the 10% level. The residual ACF and PACF are given in the following two slides. The ACF and PACF are well within their two standard error limits. Final Estimates of Parameters Type Coef SE Coef T P AR 1 -0. 3243 0. 1246 -2. 60 0. 012 AR 2 0. 2192 0. 1251 1. 75 0. 085 Constant 284. 903 6. 573 43. 34 0. 000

Example

Example n n The p-value for the Ljung-Box statistics for m = 12, 24, 36, and 48 are all large (> 5%) indicating an adequate model. We use the model to generate forecasts for periods 66 and 67. MS = 2808 DF = 62 Modified Box-Pierce (Ljung-Box) Chi. Square statistic Lag 12 24 36 48 Chi-Square 6. 3 13. 3 18. 2 29. 1 DF 9 21 33 45 P-Value 0. 707 0. 899 0. 983 0. 969

Example n The forecasts are generated by the following equation.

Example n n The 95% prediction limits are approximately The 95% prediction limits for period 66 are

Final comments n n In ARIMA modeling, it is not good practice to include AR and MA parameters to “cover all possibilities” suggested by the sample ACF and Sample PACF. This means, when in doubt, start with a model containing few parameters rather than many parameters. The need for additional parameters will be evident from the residual ACF and PACF.

Final comments n Least square estimates of AR and MA parameters in ARIMA models tend to be highly correlated. When there are more parameters than necessary, this leads to unstable models that can produce poor forecasts.

Final comments n n To summarize, start with a small number of clearly justifiable parameters and add one parameter at a time as needed. If parameters in a fitted ARIMA model are not significant, delete one parameter at a time and refit the model. Because of high correlation among estimated parameters, it may be the case that a previously non-significant parameter becomes significant.